A Study on Rephrasing Large Earthquakes via the Least Action Principle and Evolving Machine Learning
Abstract
Understanding large earthquakes (EQs) has been a critical endeavor. Recent advancements in machine learning (ML) help this endeavor, yet they face formidable performance ceilings due to our technical limitations, restricted access to seismogenic fault zones, and the lack of large datasets for ML training. This paper introduces a novel approach by conceptualizing the EQ-bearing lithosphere as a chaotic system and investigating whether the least action principle (LAP) from Hamiltonian mechanics can illuminate the potential predictability of large EQ initiation. Leveraging an abstract space defined by time and rupture derivatives, this paper establishes a potential causal pathway between novel LAP-based features and large EQs. An evolving ML framework utilizing parallel evolution strategies was developed to identify hidden rules derived from LAP. The model was trained on data from all EQs in the western U.S. over the past 40 years. Feasibility tests on historical large EQs show promising performance compared to state-of-the-art forecasting methods. While practical applicability requires deeper investigation, these findings suggest a new research direction for large EQ prediction that merges classical physics and evolving ML.
Publication Title
IEEE Access
Recommended Citation
Cho, I., & Choi, E. (2026). A Study on Rephrasing Large Earthquakes via the Least Action Principle and Evolving Machine Learning. IEEE Access, 14, 73744-73764. https://doi.org/10.1109/ACCESS.2026.3693178
